def main(args):
"""
Environment used in this code is Pendulum-v0 from OpenAI gym.
States: cos(theta), sin(theta), theta_dt
Actions: Force application between -2 to 2
Reward: -(Θ^2 + 0.1*Θ_dt^2 + 0.001*action^2)
Objective: Pendulum is vertical, with 0 movement.
Initialization: Starts at a random angle, and at a random velocity.
End: After all the steps are exhausted
"""
# Initialize saver
saver = tf.train.Saver()
with tf.Session() as sess:
# Create the gym environment
env = LinearSystem(nsim=args['max_episode_len'], model_type='MIMO', x0=np.array([1.333, 4]),
u0=np.array([3, 6]), xs=np.array([3.555, 4.666]), us=np.array([5, 7]), step_size=0.2)
# Set all the random seeds for the random packages
np.random.seed(int(args['random_seed']))
tf.set_random_seed(int(args['random_seed']))
env.seed(int(args['random_seed']))
tflearn.init_graph(seed=args['random_seed'])
# Define all the state and action dimensions, and the bound of the action
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
action_bound = env.action_space.high
# Restore old model
# saver.restore(sess, args['ckpt_dir'])
# Initialize the actor and critic
actor = ActorNetwork(sess, state_dim, action_dim, action_bound,
float(args['actor_lr']), float(args['tau']),
int(args['minibatch_size']))
critic = CriticNetwork(sess, state_dim, action_dim,
float(args['critic_lr']), float(args['tau']),
float(args['gamma']),
actor.get_num_trainable_vars())
# Initialize Ornstein Uhlenbeck Noise
actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(action_dim), dt=env.step_size,
random_seed=int(args['random_seed']))
# sigma=np.array([1, 1])
# Train the Actor-Critic Model
replay_buffer, action_list = train(sess, env, args, actor, critic, actor_noise)
# Save the model
saver.save(sess, args['ckpt_dir'])
return actor, critic, env, replay_buffer, action_list
def main(args):
with tf.Session() as sess:
env = gym.make(args['env'])
np.random.seed(int(args['random_seed']))
tf.set_random_seed(int(args['random_seed']))
env.seed(int(args['random_seed']))
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
action_bound = env.action_space.high
# Ensure action bound is symmetric
assert (env.action_space.high == -env.action_space.low)
actor = ActorNetwork(sess, state_dim, action_dim, action_bound,
float(args['actor_lr']), float(args['tau']),
int(args['minibatch_size']))
critic = CriticNetwork(sess, state_dim, action_dim,
float(args['critic_lr']), float(args['tau']),
float(args['gamma']),
actor.get_num_trainable_vars())
actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(action_dim))
train(sess, env, args, actor, critic, actor_noise)